Dynamic Graph Representation Learning Based On Graph Neural Networks

Graphs also known as networks are common and important data type in the real world,such as social networks,transportation networks,protein networks and so on.Generally,experts use graph analysis techniques to mine the information in graphs and apply it in tasks such as node classification,link prediction and community detection.Graph representation learning is an effective graph analysis technique that can represent graphs into low-dimensional and real-valued dense vectors based on algorithms such as random walks,matrix factorization,and graph neural networks(GNN).However,these algorithms only focus on static graphs,ignoring the dynamics of graphs.Therefore,the research of this paper focuses on how to use GNN models for dynamic graph representation learning.This paper mainly focuses on two aspects,which are as follows: 1.How can the GNN models effectively learn the temporal information of graphs;2.How to enhance the generalization of GNN models.To address the first problem,we propose a conditional variational time-series graph autoencoder TS-CVGAE,which combines conditional variational autoencoders and GNN to aggregate graph structure,node attribute and evolving information.Firstly,a time-series based graph convolution is constructed,which uses LSTM to aggregate the evolving information of graphs.Secondly,the conditional variational auto-encoder framework is used to train the model.Finally,the model is evaluated in the link prediction task.The prediction performance of TS-CVGAE outperforms the baseline methods on all four real dynamic graph datasets.The address the second problem,we propose a pre-training framework on dynamic graph neural networks named PT-DGNN.Inspired by the idea of graph generation,a dynamic pre-training method is proposed to enable the model to capture the structural,semantic and evolving information of graphs.Firstly,a series of subgraphs are sampled with local time-aware structure from graphs by using a dynamic subgraph sampling algorithm.Secondly,two dynamic graph generation tasks,i.e.dynamic edge generation and node attribute generation,are proposed to learn dynamic node representation.Finally,the model is evaluated using link prediction.The prediction performance of PT-DGNN outperforms the baseline methods on all three real dynamic graph datasets.